Transparent and efficient fault tolerance on high-speed datacenter networking

Abstract

Hardware failures are commonplace in datacenters and typically lead to catastrophic consequences. For example, in March 2017, Amazon storage services (S3) went offline for an extended period of time due to server failures in the datacenter, and this outage caused the company to lose over $150 million. To achieve fault tolerance of those important services deployed in the datacenter, replication is the de-facto standard technique. The core of replication is to provide multiple identical copies of a service on different physical machines so that the service can continue to operate even if some of its replicas fail.

Unfortunately, despite much research and engineering effort, a replication system that is both transparent and efficient is still missing. A transparent replication system runs general applications without any modifications, and can thus relieve the application developers from the intrusive and error-prone modifications to the program’s structure and code. An efficient replication system causes reasonable performance overhead to the underlying applications.

This thesis proposes three transparent and efficient replication systems, APUS, PLOVER, and GANNET. To achieve transparency, APUS, PLOVER, and GANNET utilise the unique features on three different system levels, application level, virtual machine level, and storage level, respectively. To build efficient (high-performance) systems, high-speed RDMA networking in modern datacenters is leveraged and corresponding new protocols are designed to exploit the whole potential of this emerging new hardware.